AutoRebuttal
AutoRebuttal is an evidence-first rebuttal workflow for coding agents. It parses papers and reviews, decomposes reviewer concerns, drafts venue-aware responses, and optionally routes empirical concerns into measured experiment loops with verifiable logs, metrics, and non-fabricated rebuttal insertion.
It is built for one job: help authors turn a paper, reviews, and explicit rebuttal constraints into a structured, evidence-first response without fabricating experiments, gains, or citations.
The current repo proves three paper-input lanes:
- paper PDF
- extracted or manually supplied paper text
- LaTeX paper input as either a single
.tex file or a directory containing .tex files
Review inputs remain PDF or text. Revise mode still starts from an existing rebuttal PDF or rebuttal text. OCR support is limited to the implemented rendered-page fallback path in skills/auto-rebuttal/scripts/; this repo does not claim generic OCR or full LaTeX compilation/edit automation beyond the helpers that already exist.
The package exposes three command-style entrypoints:
/rebuttal for drafting from paper + reviews/rebuttal_revise for polishing an existing rebuttal draft/experiment-bridge for supplementary evidence when reviewers ask for new experiments
Why Evidence-First Rebuttal?
Review responses often fail in two opposite ways: they either overpromise new experiments that are not actually run, or they bury real evidence in unstructured logs that cannot be traced back to the rebuttal claim. AutoRebuttal keeps the drafting layer and evidence layer connected but separate.
The workflow can now represent a reviewer request as an Experiment Request, map it into a runnable Experiment Packet, run or materialize that packet, parse a metric, and write an Evidence Ledger entry. The ledger is the source of truth for experimental claims; rebuttal prose should only use verified claims or explicit placeholders.
What AutoRebuttal Can And Cannot Do
AutoRebuttal can:
- parse papers, reviews, and existing rebuttals into structured drafting bundles
- classify empirical reviewer concerns such as ablations, baselines, runtime, robustness, datasets, metrics, and failure cases
- run a local shell experiment packet with a timeout and captured stdout/stderr
- parse JSON, CSV, or regex-log metrics into measured values
- compare a candidate metric with a baseline value when provided
- generate a Slurm
sbatch script for CV/HPC-style dry runs
- write
results.tsv or results.jsonl plus evidence_ledger.json
AutoRebuttal cannot:
- infer the correct training or evaluation command for an arbitrary project without author mapping
- guarantee that a cluster accepts a generated Slurm script
- prove that a model improved unless a real command, log, result file, and metric parser support the claim
- safely roll back arbitrary project edits without the author's version-control policy
- insert unverified experiment results into rebuttal prose as if they were verified
AutoRebuttal Outputs
Quick Install
Codex
Tell Codex:
Fetch and follow instructions from https://raw.githubusercontent.com/YoujunZhao/AutoRebuttal/refs/heads/main/.codex/INSTALL.md
Claude Code
Install it through the Claude plugin workflow, use:
/plugin marketplace add YoujunZhao/AutoRebuttal
/plugin install auto-rebuttal@auto-rebuttal-dev
Other installation
Codex
Preferred path: native skill discovery via clone + junction / symlink.
Clone the repo:
git clone https://github.com/YoujunZhao/AutoRebuttal.git ~/.codex/AutoRebuttal
Create the skill symlink:
mkdir -p ~/.agents/skills
ln -s ~/.codex/AutoRebuttal/skills/auto-rebuttal ~/.agents/skills/auto-rebuttal
Windows (PowerShell):
New-Item -ItemType Directory -Force -Path "$env:USERPROFILE\.agents\skills"
cmd /c mklink /J "$env:USERPROFILE\.agents\skills\auto-rebuttal" "$env:USERPROFILE\.codex\AutoRebuttal\skills\auto-rebuttal"
Update through the clone:
cd ~/.codex/AutoRebuttal && git pull
Optional manager CLI path:
